Method for identifying and grading stressed pseudo-mature fruits of litsea cubeba

Abstract

The present application relates to the technical field of image data processing, in particular to a recognition and grading method for stressed pseudo-mature fruits of Litsea cubeba, comprising: collecting thermal infrared time series images and hyperspectral images of Litsea cubeba fruits, and extracting thermal lag index, pseudo-mature spectral index and glycoside index; constructing the above features, fruit diameter, surface temperature standard deviation and spatial position into a spatiotemporal graph sequence, using a spatiotemporal graph convolution network with an environmental stress memory gate to output a pseudo-mature probability and define a pseudo-mature index; and according to the value, dividing the fruits into three grades of normal, mild pseudo-mature and severe pseudo-mature, and outputting the grading results. Through the synergistic integration of thermal physical anomaly detection, spectral layer inversion and environmental stress delay memory, the present application breaks the technical prejudice that "appearance maturity is equivalent to internal quality" in the field, and realizes the nondestructive and accurate recognition and grading of "phenotype-genotype decoupling" type pseudo-mature fruits.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a method for identifying and grading stress-induced false ripening fruits of Litsea cubeba. Background Technology

[0002] The citral content in Litsea cubeba fruit is a core quality indicator determining its economic value, and precise control of the harvesting time directly determines the final quality grade of the essential oil product. Currently, the mainstream method for judging the maturity of Litsea cubeba fruit in production practice mainly relies on external phenotypic characteristics such as changes in peel color (from greenish-blue to bluish-green with white spots) and the intensity of the aroma after being crushed by hand. Based on long-term production practice, a relatively fixed cognitive model has been formed in this field, which believes that there is a basically stable correspondence between the external phenotypic characteristics of the fruit and the internal citral accumulation, and harvesting can be arranged when the peel shows a suitable appearance for harvesting.

[0003] However, a type of quality anomaly has long been overlooked in actual cultivation. When Litsea cubeba plants encounter abiotic stresses such as prolonged high temperatures and drought during the critical period of fruit growth and development, the fruit will still gradually exhibit a "harvestable" appearance according to a predetermined process—the peel turns pale green and white spots appear. However, the internal biosynthetic pathway of citral is actually inhibited, and carbon flow is redistributed to the synthesis of stress-related secondary metabolites, resulting in a severe deficiency of citral content. Some citral is also sealed inside the cells in a glycosidic-bound form. At this time, the "appearance maturity" and "internal quality" of the fruit have decoupled, forming a so-called false ripening state. Because this anomaly only involves the deviation of the internal physiological metabolic pathway of the fruit while the external morphological changes conform to the normal ripening process, it is completely impossible to distinguish it from normal harvestable fruit by existing visual observation and olfactory judgment. After being mixed in, the citral content of the entire batch of essential oil will significantly deviate from the expected quality standard, leading to product downgrading or even scrapping, causing hidden but serious economic losses to growers and processing enterprises.

[0004] The root cause of these problems lies in the fact that existing technologies are limited to acquiring and analyzing the "surface information" of the fruit, lacking effective means to perceive abnormalities in the internal tissue state and thermophysical properties of the fruit. Furthermore, there is a lack of technical approaches to link environmental stress history with the physiological response of the fruit. This results in a large number of falsely ripened fruits failing to be detected and removed in a timely manner during harvesting, and subsequent processing also failing to effectively grade them. Therefore, how to break through the existing maturity determination paradigm that relies on external phenotypic appearance and achieve non-destructive identification and grading of falsely ripened Litsea cubeba fruits is a concentrated and long-neglected technical challenge that urgently needs to be addressed in this field. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to break through the existing maturity determination paradigm that relies on appearance phenotype, and achieve non-destructive identification and grading of falsely ripe fruits of Litsea cubeba.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The identification and grading method for stress-induced false ripening fruits of Litsea cubeba includes the following steps: S1: Acquire thermal infrared time-series images and hyperspectral images of the target Litsea cubeba fruit; extract the surface temperature change sequence of each fruit from the thermal infrared time-series images; and calculate the thermal time constant. and thermal hysteresis index The thermal hysteresis index Defined as the cumulative difference between the fruit cooling rate curve and the normal reference curve, where the differential region is indicated by a threshold of the second derivative of the rate of temperature change; the epidermal reflectance is obtained from the hyperspectral image through a two-layer radiative transfer model. and the reflectivity of the pulp layer Calculate the pseudo-ripe spectral index at the characteristic wavelength. and glycoside index ; S2: The multiple fruits on the same branch... Using fruit diameter and surface temperature standard deviation as node features, and spatial adjacency relationships between fruits as edges, a spatiotemporal graph sequence is constructed. A spatiotemporal graph convolutional network is then used to extract features from this sequence. This network incorporates an environmental stress memory gating mechanism, which takes temperature, solar radiation, and soil moisture data from the past 72 hours as input and modulates the graph convolution output, ultimately outputting the false ripening probability of each fruit. ; S3: Define the false maturity index and according to the The value is used to grade the fruit: when At that time, it was judged to be a normal fruit. At that time, it was judged to be a lightly ripened, semi-ripe fruit. It was judged to be a heavily overripe fruit; S4: Output the grading results for each fruit, which will be used as the basis for subsequent selective processing decisions.

[0007] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripening fruits of Litsea cubeba, the thermal hysteresis index mentioned in S1... The calculation formula is: in, The number of time sampling points in this invention ; For the fruit at time Excess temperature; For the normal reference curve at time The excessive temperature, the normal reference curve is established from normal fruit samples that have not been subjected to stress; The curvature threshold of the second derivative; The first derivative of excess temperature with respect to time at time... The value; Number of time sampling points; I This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0008] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripening fruits of Litsea cubeba, the false-ripening spectral index mentioned in S1... The calculation formula is: in, The reflectance of the pulp layer at a wavelength of 1200 nm; The reflectance of the pulp layer at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1200 nm; The glycoside index Defined as the ratio of the peak height of the second derivative spectrum to the baseline for hyperspectral reflectance in the range of 1420–1480 nm.

[0009] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the total reflectance of the double-layer radiative transfer model described in S1 is expressed as: in, Wavelength variable (unit: nm); The reflectance of the epidermis; The transmittance of the epidermis represents the proportion of energy remaining after light passes through the epidermis once. The reflectance of the pulp layer; The absorption and scattering coefficients of the epidermis and pulp layers were fitted using a full-spectrum global parameterization method. The set of global parameters to be fitted is as follows: , , , , , , , , , in, , , These are the concentration weighting coefficients for water, lipids, and citral in the epidermis, respectively. , , These are the concentration weighting coefficients for water, lipids, and citral in the pulp layer, respectively. and These are the scattering coefficients of the epidermis and pulp layer at a wavelength of 1000 nm, respectively. and These are the scattering power indices of the epidermis and pulp, respectively.

[0010] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the absorption coefficients of the epidermis and pulp are respectively expressed as linear combinations of the absorption basis functions of the three chemical components: in, The absorption coefficient of the epidermis; The absorption coefficient of the pulp layer; , , These are the basis functions for the mass absorption coefficients of water, lipids, and citral, respectively. , , , respectively, are the concentration weighting coefficients (unit: g / L) of water, lipids and citral in the epidermis, and are the global parameters to be fitted; , , , respectively, are the concentration weighting coefficients (unit: g / L) of water, lipid, and citral in the pulp layer, and are the global parameters to be fitted; The scattering coefficients of the epidermis and pulp layers are approximated using the power-law approximation of Mie scattering. in, and These are the scattering coefficients of the epidermis and pulp at a reference wavelength of 1000 nm, respectively, and are the global parameters to be fitted. and These are the scattering power indices of the epidermis and pulp, respectively.

[0011] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the spatial convolution and temporal convolution of the spatiotemporal graph convolutional network described in S2 are respectively defined as: in, The node currently being processed; This is the index for the current time slice; For nodes The neighborhood set; These are the normalization coefficients; For nodes In time slice eigenvectors; For nodes In the previous time slot eigenvectors; The weight matrix is ​​the spatial convolution kernel. This is the temporal convolution kernel weight matrix; For nodes In time slice The output feature vector after spatial graph convolution; The implementation method of the environmental stress memory gating is as follows: in, The feature vector output by the spatial graph convolutional layer; The environmental feature vector is represented by principal component scores for each element. To be and The joint vector obtained by concatenation; This is a gated weight matrix, where each element is a weight. This is a bias term, and each element is a bias. It is the Sigmoid activation function. This reduces the gate value to between 0 and 1. This is a gating vector, with each element taking values ​​from 0 to 1, controlling the mixing ratio of new and old features; For Hadamard product; This is the final output feature vector of this node in the previous time slice; This is the final output feature vector after gating modulation, where each element is an activation value.

[0012] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the method for determining the spatial adjacency relationship described in S2 is: the distance between any two fruits. Calculated from three-dimensional spatial coordinates: in, and Fruits and fruit Three-dimensional spatial coordinates; For fruit and fruit The Euclidean distance between them.

[0013] Edge weight The Gaussian kernel function defined as distance: in, This refers to the spatial scale parameter.

[0014] Furthermore, in the aforementioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the spatiotemporal graph convolutional network comprises three layers of spatiotemporal graph convolutional modules, with each layer outputting the hidden layer feature dimension. The probability of false maturity The 64-dimensional feature vector output from the third-layer spatiotemporal graph convolution is mapped to a 2-dimensional logical value through a fully connected layer, and then normalized by the Softmax function.

[0015] Furthermore, in the above-mentioned method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba, the training of the spatiotemporal graph convolutional network adopts a two-stage strategy: the first stage uses labeled source domain data for supervised training, employing a binary cross-entropy loss function. in, For the first The true label of each sample (1 indicates false maturity, 0 indicates normal); The model predicts the first The probability of false maturity for each sample; This represents the total number of labeled samples. This represents the cross-entropy loss value. The second stage uses CORAL domain adaptation to fit the unlabeled target domain data, with the loss function being: in, Let be the source domain feature covariance matrix, whose elements are the covariance of the source domain samples on the output features of the third-layer spatiotemporal graph convolution; Let be the target domain feature covariance matrix, whose elements are the covariance of the target domain samples on the output features of the third-layer spatiotemporal graph convolution; For feature dimensions; It is the Frobenius norm.

[0016] The beneficial effects of this invention are as follows: By introducing thermal inertia analysis theory, this invention transforms the abnormal fruit cooling rate contained in thermal infrared time-series images into a quantitatively calculable thermal hysteresis index. Simultaneously, it introduces a layered radiative transfer model from the field of remote sensing optics, and for the first time achieves precise separation of the spectral characteristics of the epidermis and pulp of *Litsea cubeba* through a full-spectrum global parameterized inversion strategy. This allows for the construction of a false-ripening spectral index and a glycoside index, respectively identifying two complementary false-ripening indicators: abnormal thickening of epidermal wax and accumulation of citral in the pulp glycoside-bound state. This forms a non-destructive detection system that collaboratively identifies false-ripening fruits from two independent dimensions: "abnormal thermal response" and "spectral layered distortion." Furthermore, this invention introduces the delayed memory effect of environmental stress into a deep learning architecture, enabling the model not only to perceive the current spectral and thermal state of the fruit but also to "remember" the time-cumulative effect after the stress event, overcoming the limitation of existing static detection models that can only reflect instantaneous states. This invention solves the fundamental problem that existing technologies cannot identify stress-induced false ripening fruits, and changes the technical paradigm of relying on appearance experience in the harvesting of Litsea cubeba. It enables false ripening fruits that were originally mixed into normal batches, causing the quality of the entire batch of essential oil to be downgraded, to be identified one by one and graded and sorted, providing a decision-making basis for subsequent selective processing. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for identifying and grading stress-induced false ripening fruits of Litsea cubeba, as described in a specific embodiment of the present invention. Detailed Implementation

[0018] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0019] Reference Figure 1 This invention relates to a method for identifying and grading stress-induced false ripening fruits of Litsea cubeba, comprising the following steps: S1: Acquire thermal infrared time-series images and hyperspectral images of the target Litsea cubeba fruit; extract the surface temperature change sequence of each fruit from the thermal infrared time-series images; and calculate the thermal time constant. and thermal hysteresis index The thermal hysteresis index Defined as the cumulative difference between the fruit cooling rate curve and the normal reference curve, where the differential region is indicated by a threshold of the second derivative of the rate of temperature change; the epidermal reflectance is obtained from the hyperspectral image through a two-layer radiative transfer model. and the reflectivity of the pulp layer Calculate the pseudo-ripe spectral index at the characteristic wavelength. and glycoside index ; In this embodiment, the experimental site was located at the Litsea cubeba planting base in Badu Town, Jiaocheng District, Ningde City, Fujian Province (26°65′N, 119°52′E, altitude 280 m), and the variety was the local "Funing Daguo" Litsea cubeba. This region has a mid-subtropical maritime monsoon climate with an average annual temperature of 18.6℃ and an average annual rainfall of approximately 2100 mm. The soil is mountain yellow-red soil. From July 15th to July 24th, 2024, the base experienced a prolonged high-temperature drought stress event, during which the daily maximum temperature reached 35.8–39.2℃, the cumulative rainfall was only 3.5 mm, and the soil volumetric moisture content dropped to 35%–40%.

[0020] Thirty-two healthy 5-8 year old Litsea cubeba trees were selected from the most severely stressed southern slope (approximately 20° slope, sandy yellow-red soil) as experimental subjects. Simultaneously, 12 unstressed trees of the same variety and similar age were selected from a shaded area on the northern slope within the same site as a normal control group. No additional agricultural operations such as irrigation or fertilization were performed on any of the experimental plants during the experiment.

[0021] Sampling was conducted on the second day (July 26), the fourth day (July 28), the sixth day (July 30), and the eighth day (August 1) after the end of the stress event. At each sampling, five intact fruits without mechanical damage were selected from each of the south-facing, southeast-facing, and southwest-facing branches of each stressed plant, for a total of 20 plants × 3 orientations × 5 fruits = 300 fruits / sampling, and a total of 1200 samples across the four samplings. The control group of 12 plants was sampled using the same method, with 240 fruits collected each time, for a total of two samplings (July 26 and July 30).

[0022] Each fruit sample was immediately numbered, weighed, and recorded after thermal infrared and hyperspectral data acquisition. Some samples were used for subsequent GC-MS benchmark measurements to verify the model prediction results.

[0023] A FLIR A655sc long-wave infrared thermal imager (spectral response range 7.5–14 μm, infrared image resolution 640 × 512 pixels) was used to acquire time-series thermal infrared images. The thermal imager was fixed on a tripod, with the lens axis perpendicular to the equatorial plane of the fruit, and the shooting distance was 0.5 m. A small piece of black tape (5 mm × 5 mm, emissivity) was attached to the surface of each fruit to be tested. ), used for temperature reference calibration.

[0024] During collection, carefully pluck the fruit from the branch (avoiding damage to the peel as much as possible) and immediately place it on an insulated foam support in the center of the thermal imager's field of view (the contact area between the support and the fruit should be minimal to reduce heat conduction interference). Start the thermal imager in continuous shooting mode, sampling at a frequency of 1 Hz, and continuously shoot for 120 seconds, acquiring a total of 120 frames of thermal infrared time-series images. Simultaneously record the ambient temperature. (Unit: °C) and wind speed (Unit: m / s) Measurements were taken using the built-in miniature meteorological sensor of the thermal imager and an external anemometer (accuracy ±0.1 m / s).

[0025] For each frame of the thermal infrared image, the Otsu adaptive thresholding method is first used for fruit region segmentation. Specifically, for the first frame... Frame image ( (For frame index), calculate the image grayscale histogram, and find the threshold that maximizes the inter-class variance. ( (Grayscale threshold) in, The grayscale segmentation threshold (value range 0~255); and These represent the proportions of the background and target regions to the total number of pixels in the entire image (between 0 and 1). and These are the average gray values ​​(between 0 and 255) of the background and target regions, respectively. The variance is between classes.

[0026] exist Range search The largest Value, denoted as With threshold The image is binarized and segmented to obtain a binary mask of the fruit region. Then, a morphological opening operation is performed using a circular structuring element with a radius of 3 pixels to eliminate isolated noise and smooth the boundaries of the fruit region.

[0027] The segmented fruit region is mapped back to the radiation temperature data matrix of the original thermal infrared image. The radiation temperature values ​​of all pixels within the region are extracted, and the arithmetic mean is taken as the average surface temperature of the fruit in that frame image. (Unit: °C). The temperature value of the black tape area was extracted simultaneously as an emissivity calibration reference (the radiation temperature correction factor of the black tape is known).

[0028] For each fruit, the average surface temperature over 120 frames was measured. With the corresponding ambient temperature By performing a difference operation, we obtain the excess temperature sequence: in, For the first The acquisition time corresponding to the frame (unit: seconds). s, s; For a moment Average surface temperature of the fruit (unit: °C); For a moment Ambient temperature (unit: °C); For a moment Excess temperature (unit: °C) represents the excess temperature of the fruit surface relative to the ambient air.

[0029] Excess temperature sequence for each fruit A nonlinear least-squares fit was performed using an exponential decay model based on Newton's law of cooling. Theoretically, the cooling process of the fruit surface after the heat source is removed satisfies: in, For time variables (unit: seconds); for Excess temperature at any given time (unit: °C); The initial excess temperature, i.e. time Value (unit: °C); The thermal time constant (unit: seconds) is the core parameter to be fitted; The random error term (unit: °C) follows a normal distribution with a mean of zero.

[0030] The Levenberg-Marquardt algorithm is used to find the parameters that minimize the sum of squared residuals. and : The algorithm's initial value is set to , The convergence condition for iteration is set to the parameter change being less than 10. -6 Or the number of iterations exceeds 100. Record the results after fitting is complete. The estimated value (unit: s).

[0031] Based on the thermal infrared data of 240 normal fruit samples collected on July 26th from the control group (unstressed plants in the north-facing shaded area), a standard cooling curve for normal fruits was established. The arithmetic mean of the excess temperature sequence for each of the 240 normal fruits was calculated: in, Index for normal fruit numbers ( ); For the first A normal fruit at any time Excess temperature value (unit: °C); This represents the total number of samples from normal fruits. ; For the normal reference curve at time The average excess temperature (unit: °C).

[0032] For each fruit to be tested, calculate the thermal hysteresis index according to the formula. : in, The number of time sampling points in this invention ; For the fruit at time Excess temperature (unit: °C); For the normal reference curve at time Excess temperature (unit: °C); The first derivative of excess temperature with respect to time at time... The value (unit: ℃ / s) is calculated using the central difference method: For endpoints and Forward difference and backward difference are used respectively; For the normal reference curve at time The first derivative value (unit: ℃ / s) is calculated using the method described above. Completely consistent, that is, for The central difference method (using forward / backward difference at the endpoints) is also used for numerical differentiation; The second derivative of excess temperature with respect to time at time... The value (unit: ℃ / s²) is also calculated using the central difference method; The curvature threshold of the second derivative was determined experimentally. ℃ / s² (unit: ℃ / s²); This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0033] The physical meaning of this formula is: only when the second derivative of the cooling curve exceeds the threshold... Only at the moment when the curve shows a clear "inflection point" or "plateau" is the difference between the current fruit cooling rate and the cooling rate of the normal reference curve accumulated. . The unit is ℃ / s (degrees Celsius per second). The higher the value, the more severe the deviation of the fruit's cooling behavior from the normal pattern, and the higher the possibility of false ripening.

[0034] Hyperspectral images were acquired using a Finnish Specim FX17 hyperspectral camera (wavelength range 900–1700 nm, spectral resolution 8 nm, spatial resolution 640 pixels). The camera was mounted on a linear translation stage, employing a line-scan pushbroom imaging method. Acquisition was conducted in a darkroom (ambient temperature 22±1℃, relative humidity 50±5%) to eliminate stray light interference. Hyperspectral images were acquired for each fruit in four directions along the equatorial plane. Specifically, the fruit was placed on the rotating clamp of the hyperspectral camera stage, ensuring the fruit's equatorial plane was perpendicular to the camera's optical axis and the fruit's center was at the center of the field of view. The clamp was rotated sequentially to 0°, 90°, 180°, and 270° (rotation angle tolerance ±1°), acquiring one hyperspectral image for each position. Each image contained 640×400 pixels of spatial information and 224 spectral bands. After acquisition, the arithmetic mean of the reflectance spectral data acquired in the four directions was calculated for each wavelength point, serving as the representative spectrum of the fruit. To eliminate the influence of fruit surface non-uniformity (such as local light differences and uneven distribution of surface wax) on spectral measurements.

[0035] Before acquisition, white field correction was performed using a standard white board (Spectralon® material with 99% reflectivity). Dark current images were acquired with the camera shutter closed for dark field correction. The corrected reflectivity... Calculate using the following formula: in, The wavelength variable (unit: nm) is taken in intervals of 8 nm within the range of 900~1700 nm; For the sample at wavelength The original DN value (digital signal) at the location; Dark field DN value, which is the camera's background signal when the shutter is closed; For standard whiteboards at wavelength The DN value at that location.

[0036] Original reflectance spectrum of each fruit The following preprocessing steps are performed sequentially: Savitzky-Golay convolution smoothing: An SG filter with a window width of 11 wavelengths and a polynomial order of 3 is used for smoothing to remove high-frequency noise and obtain the smoothed spectrum. (Reflectivity); Standard Normal Transform (SNV): For smoothed spectra Standardization process: in, The smoothed reflectance (between 0 and 1); for The arithmetic mean over the entire wavelength range (900~1700 nm, approximately 100 effective wavelength points); for Standard deviation across the entire spectral range; The standardized reflectance has a mean of 0 and a standard deviation of 1.

[0037] The fruit of *Litsea cubeba* can be considered a two-layered, turbid medium consisting of an epidermis (including a waxy layer) and a pulp layer. Total reflectance. The sum of the contribution to reflection from the epidermis and the contribution to reflection returning from the pulp layer after transmission through the epidermis: in, Wavelength variable (unit: nm); The reflectance of the epidermis (between 0 and 1); The transmittance of the epidermis (between 0 and 1) represents the proportion of energy remaining after light passes through the epidermis once. "Indicates the total transmittance of light that passes through the epidermis once upon incident, is reflected from the pulp layer, and then passes through the epidermis once more." The reflectance of the pulp layer (between 0 and 1).

[0038] In this embodiment, to avoid the underdetermined problem caused by independent single-wavelength inversion, a full-spectrum global parameterization inversion strategy is adopted. Specifically, the absorption coefficient is... and scattering coefficient The parameters are parameterized into a linear combination and power-law form of known basis functions. Using reflectance data from all approximately 100 effective wavelength points in the range of 900–1700 nm, the global parameters are fitted in one go.

[0039] The optical behavior of the epidermis and pulp layers is described using the Kubelka-Munk two-stream approximation theory. For the epidermis, its absorption coefficient... (Unit: mm) -1 ) and scattering coefficient (Unit: mm) -1 ) and reflectivity and transmittance The following relationship exists: in: Absorption coefficient of the epidermis (unit: mm) -1 This indicates the degree of energy attenuation of light when it travels a unit distance in the epidermis; Scattering coefficient of the epidermis (unit: mm) -1 ), representing the probability that light is scattered when it travels a unit distance in the epidermis; It is the first auxiliary variable in the Kubelka-Munk theory and is always greater than 1; It is the second auxiliary variable in the Kubelka-Munk theory and is always greater than 0; This is an auxiliary variable for the interface reflection coefficient, ranging from 0 to 1; The effective thickness of the epidermis (unit: mm) is taken as an empirical value for Litsea cubeba fruit. ; It is a hyperbolic cotangent function. It is a hyperbolic sine function. These are hyperbolic cosine functions, both of which are mathematical functions.

[0040] The pulp layer can be considered as a thick optical medium (its thickness is much greater than the mean free path of a photon), and its reflectivity... Diffuse reflectivity of an approximately semi-infinite medium: in, Absorption coefficient of the pulp layer (unit: mm) -1 ); Scattering coefficient of the pulp layer (unit: mm) -1 ); The internal reflection coefficient at the interface between the pulp layer and the epidermis layer is taken as an empirical value. .

[0041] To be based on actual measurements China's stable reverse deduction and ,right and The following global parameterization is performed; the absorption coefficient is composed of a linear combination of the absorption basis functions of the three main chemical components: in, , , The basis functions of known mass absorption coefficients (in L·g) for water, lipids (represented by waxes and cell membrane lipids), and citral in the 900–1700 nm wavelength range are given. -1 ·mm -1 The aforementioned basis functions are common knowledge in this field and can be obtained from the NIST (National Institute of Standards and Technology) Standard Reference Database (SRD 203a) or published near-infrared spectroscopy literature (such as Workman & Weyer, Practical Guide to Interpretive Near-Infrared Spectroscopy, CRC Press, 2007). Water exhibits characteristic absorptions at the first overtone absorption peak of the OH stretching vibration near 1450 nm and the combination absorption peak near 1940 nm (typical mass absorption coefficients are approximately 0.8–1.2 L·g, respectively). -1 ·mm -1 and 1.0~1.5 L·g -1 ·mm -1 Lipids exhibit characteristic absorption at the first overtone of the CH stretching vibration near 1730 nm (typical mass absorption coefficient is approximately 0.5–0.8 L·g). -1 ·mm -1 Citral exhibits characteristic absorption at the CH second ohm band near 1200 nm (typical mass absorption coefficient is approximately 0.3–0.6 L·g).-1 ·mm -1 In the algorithm implementation, these basis functions are input as a numerical table at discrete wavelength points, with a wavelength step size consistent with the hyperspectral imager (8 nm).

[0042] , , , respectively, are the concentration weighting coefficients (unit: g / L) of water, lipids and citral in the epidermis, and are the global parameters to be fitted; , , , , represent the concentration weighting coefficients (unit: g / L) of water, lipid, and citral in the pulp layer, respectively, and , , represent the global parameters to be fitted.

[0043] The scattering coefficients are approximated using the power-law approximation of Mie scattering: in, and The scattering coefficients of the epidermis and pulp at a reference wavelength of 1000 nm are shown in mm. -1 ), which are the global parameters to be fitted; and , respectively, represent the scattering power exponents of the epidermis and pulp layer, and , are the global parameters to be fitted, with typical values ​​between 0.5 and 2.0.

[0044] In this embodiment, the second auxiliary variable of the Kubelka-Munk theory is uniformly used. (with wavelength parameter) , indicating that it is a function of wavelength), while the scattering power exponent is uniformly used. and (Without wavelength parameter and with subscript), the two are completely different in symbol form and there is no conflict.

[0045] The set of global parameters to be fitted in the full-spectrum inversion is: , , , , , , , , , There are 10 parameters in total. The objective function is the sum of squares of the differences between the measured total reflectance and the model-predicted total reflectance across the entire wavelength range (900~1700 nm, approximately 100 effective wavelength points): in, It is a set of all effective wavelength points in the range of 900~1700 nm, about 100 wavelength points (according to the actual sampling interval of the spectrometer, the interval between adjacent wavelength points is 8 nm, and there are about 100 usable points in the range of 900~1700 nm). The measured total reflectance, i.e., the corrected total reflectance. Smoothed ; To set the parameters The theoretical total reflectance calculated after substituting into the above double-layer Kubelka-Munk model is, i.e. ,in , and All are performed according to the formulas described above in this step. Calculated.

[0046] This inversion problem is an overdetermined system with 10 unknowns and approximately 100 observation equations, possessing a unique stable solution. The trust-region reflective algorithm is used to solve it, with the initial values ​​set as follows: all... The coefficient is taken as 0.5 g / L. mm -1 , mm -1 , The convergence condition for the iteration is set to the objective function changing by less than 10. -8 Or the number of iterations exceeds 500. After convergence, the optimal parameters are... Substituting into the Kubelka-Munk forward modeling formula, the wavelength at each wavelength can be stably calculated. and The only solution.

[0047] After completing the spectral layer inversion, the epidermal reflectance at the following two characteristic wavelengths was extracted. and the reflectivity of the pulp layer : 1200 nm: This wavelength is near the second harmonic absorption region of CH bonds, mainly reflecting the content of terpenoid components such as citral; 1730 nm: This wavelength is near the first-order overtone absorption region of CH bond stretching vibration, which mainly reflects the density of lipids (including epidermal waxes).

[0048] False ripening spectral index Calculate using the following formula: in, The reflectance of the pulp layer at a wavelength of 1200 nm; The reflectance of the pulp layer at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1200 nm; The false ripening spectral index indicates that in normal fruit, the pulp layer has a high reflectance at 1200 nm (positively correlated with citral content) and a moderate reflectance at 1730 nm (correlated with lipid density); the epidermis layer has a low reflectance at 1730 nm (reflecting wax thickness). Therefore, normal fruit... Approaching 1.0. In falsely ripened fruits, the thickening of the waxy epidermis makes... Increased, reduced citral in fruit pulp The decrease, the combined effect of the two effects, makes Significantly greater than 1.0.

[0049] The second derivative spectra of the preprocessed spectra were calculated in the range of 1420–1480 nm. The numerical calculation of the second derivative employed the Savitzky-Golay second derivative method (window width of 9 wavelength points, polynomial order 3). Specifically, for each wavelength point… ( (where the wavelength is the index), its second derivative is approximately: in, The spectral sampling interval, nm; For the first Wavelength values ​​at each wavelength point (unit: nm); These are the Savitzky-Golay second derivative convolution coefficients. The corresponding coefficients are as follows: ; Wavelength point Standardized reflectance at that location; Wavelength point The second derivative value at (nm) -2 ).

[0050] Searching for second derivative spectra in the 1420–1480 nm range. The minimum point (this region is the absorption band, where the second derivative is negative, and the point with the largest absolute value is the location of the strongest absorption peak) is recorded, along with the corresponding wavelength. (Unit: nm) and minimum height (Absolute value of the second derivative, unit: nm) -2 Simultaneously, using the line connecting the second derivative values ​​at 1420 nm and 1480 nm as the baseline, the baseline value at the minimum point is calculated. Unit: nm -2 ), then the glycoside index Defined as: in, The absolute value of the second derivative at the local minimum point (unit: nm) -2 ); The second derivative of the baseline at the minimum point (unit: nm) -2 ); For a very small positive number (take) Unit: nm -2 To prevent the denominator from being zero; The value is the glycoside index. The higher the value, the more obvious the absorption characteristics in the 1420~1480 nm region, that is, the higher the content of glycoside-bound citral.

[0051] S2: The multiple fruits on the same branch... Using fruit diameter and surface temperature standard deviation as node features, and spatial adjacency relationships between fruits as edges, a spatiotemporal graph sequence is constructed. A spatiotemporal graph convolutional network is then used to extract features from this sequence. This network incorporates an environmental stress memory gating mechanism, which takes temperature, solar radiation, and soil moisture data from the past 72 hours as input and modulates the graph convolution output, ultimately outputting the false ripening probability of each fruit. ; In this embodiment, each sampled fruit is considered as a node in the spatiotemporal graph. The feature vector of each node It includes the following 6 dimensions: Thermal hysteresis index (unit: ℃ / s); The false ripening spectral index; Glycoside index; The standard deviation of fruit surface temperature (unit: °C) is the average surface temperature of the fruit in 120 frames of thermal infrared images. Standard deviation; The diameter of the fruit (in mm) is calculated from the spatial resolution of the hyperspectral image using a scale bar. Fruit spatial coordinates Determined by the relative position on the branch, used for subsequent determination of spatial adjacency relationships (unit: cm).

[0052] The distance between any two fruits on the same branch (Unit: cm) Calculated from three-dimensional spatial coordinates: in, and Fruits and fruit Three-dimensional spatial coordinates (unit: cm); For fruit and fruit The Euclidean distance between them (unit: cm).

[0053] like If cm, then a spatial adjacency relationship exists between the two fruits, and an undirected edge is constructed in the graph. Edge weights The Gaussian kernel function defined as distance: in, For spatial scale parameters, take cm (unit: cm); This represents the edge weight, with values ​​between 0 and 1. Fruits that are closer to each other have a higher weight.

[0054] The four samplings (days 2, 4, 6, and 8 after the end of the stress) were each constructed into four time slices. ( (For time-slice indexing). Within each time slice, all sampled fruits form a static graph. ; in, This is the set of nodes within this time slice (each node corresponds to a fruit). It is a set of edges (each edge corresponds to a pair of spatially adjacent fruits); The corresponding edge weight matrix (dimension 1) Each element is ).

[0055] The same node is connected across different time slices via time edges (i.e., the same fruit in...). and (Establish temporal correlations between feature vectors at different times) to form a complete spatiotemporal graph sequence. .

[0056] A spatiotemporal graph convolutional network based on the PyTorch Geometric framework is used. The network consists of three stacked spatiotemporal graph convolutional modules, with each layer outputting the hidden layer feature dimension. Each module contains: A spatial graph convolutional layer: performs graph convolution on the spatial neighborhood within each time slice; A temporal convolutional layer: performs one-dimensional convolution on the historical features of the same node along the time dimension.

[0057] The neighborhood of a spatial graph convolution is defined as a first-order neighborhood (i.e., directly connected nodes plus the node itself), and the normalization coefficients are... Equal to node The number of neighboring nodes ( The output of the spatial graph convolutional layer is: in, The node (fruit) currently being processed; index of the current time slice ( ); For nodes The neighborhood set (including the node itself and all nodes connected to it) (Nodes connected by edges). The normalization coefficient is equal to the node. Total number of neighboring nodes This ensures that the convolution result is not affected by the size of the neighborhood; For nodes In time slice Feature vectors (dimensions) ), whose elements are respectively , , , , In actual calculations, all features are first standardized using Z-score (i.e., the mean is subtracted and the standard deviation is divided to make the scale of each feature consistent). For nodes In the previous time slot Feature vectors (dimensions) ),when hour, Initialize with the mean vector of the source domain data (the average value of each feature dimension); The spatial convolution kernel weight matrix (learnable parameters, dimension 1) ,in For the hidden layer feature dimension, (Input feature dimension). The temporal convolution kernel weight matrix (learnable parameters, dimension 1) ,in For the hidden layer feature dimension, (Input feature dimension). For nodes In time slice The output feature vector (dimension) after spatial graph convolution Each element is an activation value.

[0058] After the spatial graph convolution output, an environmental stress memory gating module is introduced to modulate the graph convolution output using meteorological data from the past 72 hours.

[0059] First, for each fruit node Obtain hourly meteorological data for its location over the past 72 hours, including: temperature. (Unit: °C) Solar radiation Unit: W / m 2 ) and soil volumetric water content (Unit: %), of which This is an hourly index, representing hourly data for the 72 hours preceding the end of the coercive event. The original meteorological data was reduced to 8 dimensions using principal component analysis (PCA) to obtain environmental feature vectors. (Principal component score). The number of principal components retained in PCA is determined by a cumulative contribution rate ≥ 90%.

[0060] The gating unit is calculated as follows: in, The feature vector (dimension) output by the spatial graph convolutional layer Each element is an activation value; The environmental feature vector (dimension 8) is represented by principal component scores for each element. To be and The joint vector obtained by concatenation (dimension) ); For the gated weight matrix (learnable parameters, dimensions) ,Right now Each element is a weight; For bias terms (learnable parameters, dimensions) Each element is biased; It is the Sigmoid activation function. This reduces the gate value to between 0 and 1. Gated vector (dimension) Each element takes a value from 0 to 1, controlling the mixing ratio of new and old features; For Hadamard product (element-wise product); The final output feature vector (dimension) of this node in the previous time slice. ),when When, initialize with the zero vector (i.e. ); The final output feature vector (dimension) after gating modulation Each element is an activation value, which serves as the input for the next layer of spatiotemporal graph convolution.

[0061] The physical meaning of this gating mechanism: When environmental stress is severe (high temperature, drought), the gating vector... Approaching 1, the model relies more on current spectral and thermal characteristics for judgment; when environmental stress has been alleviated but the fruit response is delayed, the gating vector... As the value approaches zero, the model relies more on historical features from the previous time slice, thus "remembering" past coercive information.

[0062] After passing through 3 layers of spatiotemporal graph convolutional modules (each layer has an output dimension of 1), Each node obtains a 64-dimensional feature vector. This vector is then input into a fully connected layer ( The Softmax function then outputs the probability that the fruit belongs to either the "normal" or "false ripe" category: in, The nodes after 3 layers of spatiotemporal graph convolution The final feature vector (dimension) Each element is an activation value; The weight matrix of the fully connected layer (dimensions) Each element is a learnable parameter (weight); This represents the bias term for the fully connected layer (dimension 2), where each element is a learnable parameter (bias). The output of the fully connected layer yields a 2D logic value vector. ; Softmax is the normalized exponential function, defined as: in The first output of the fully connected layer Logical values ​​( ), It is the base of the natural logarithm (approximately 2.71828). The summation variable in the denominator summation formula is a dummy summation variable. Similarly, the value can be 1 to 2, and it is used to iterate through all categories of logical values ​​in the denominator and sum them up.

[0063] and These represent the probabilities (between 0 and 1) that the fruit belongs to the normal and false-ripe categories, respectively. .

[0064] This is the probability of the fruit being falsely ripe.

[0065] The model training dataset consists of two parts: Source Domain Data: 860 samples collected at this base in the same period of 2023 (also subjected to high temperature stress). Each sample contains complete six-dimensional node features and normal / false ripening labels determined by GC-MS (label judgment criteria: citral content <50% and glycoside ratio >25% are false ripening, otherwise normal). Target domain data: 1200 unlabeled samples collected in 2024 for unsupervised domain adaptation.

[0066] The training employs a two-stage strategy: Phase 1 (Source Domain Supervised Training): Using only 860 labeled samples, the binary cross-entropy loss function is employed. in, For sample index ( ); For the first The true label of each sample (1 indicates false maturity, 0 indicates normal); The model predicts the first The false positive probability of each sample (between 0 and 1); The total number of labeled samples. ; The natural logarithm function (in) (as base) This represents the cross-entropy loss value.

[0067] The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and 60 epochs of training. The training was performed every 10 epochs on the validation set (15% of the samples were randomly selected from 860 samples). Early stopping (patience=10, i.e., training was stopped when the validation set loss did not decrease for 10 consecutive epochs) was used to prevent overfitting.

[0068] The second stage (unsupervised adaptation to the target domain): The trained model is adapted to the CORAL domain on 1200 unlabeled samples in 2024 to minimize the covariance difference between the feature distributions of the source and target domains. in, The source domain feature covariance matrix (dimension) ,Right now ), whose elements are the covariance of the source domain samples on the output features of the third-layer spatiotemporal graph convolution; The feature covariance matrix of the target domain (dimensions) ,Right now ), whose elements are the covariance of the target domain samples on the output features of the third-layer spatiotemporal graph convolution; For feature dimension, ; For the Frobenius norm, for a matrix Its Frobenius norm is defined as ; This is the CORAL loss value.

[0069] The adaptation training used the same Adam optimizer, with the learning rate decaying to 0.0001, and trained for 20 epochs.

[0070] S3: Define the false maturity index and according to the The value is used to grade the fruit: when At that time, it was judged to be a normal fruit. At that time, it was judged to be a lightly ripened, semi-ripe fruit. The fruit was judged to be overripe and ripe.

[0071] S4: Output the grading results for each fruit, which will be used as the basis for subsequent selective processing decisions.

[0072] On July 30, 2024, 295 samples (F7.30 group) were collected. The false ripening criteria were determined using GC-MS measured citral content and glycoside percentage: citral content <50% and glycoside percentage >25%. The predictive performance of the ST-GCN model was validated. To eliminate the randomness of a single partition, a five-fold cross-validation strategy was adopted. The 295 samples were randomly divided into five equal subsets. Four subsets were used as the training set and one subset as the test set in each iteration, repeated five times and the average was taken. The number of samples correctly identified by the model in the five validations were 273, 269, 272, 270, and 271, respectively, with overall accuracies of 92.5%, 91.2%, 92.2%, 91.5%, and 91.9%, and an average accuracy of 91.9% ± 0.5% (accuracy is defined as the number of correctly identified samples divided by the total number of samples, multiplied by 100%). The total number of true positives (model-predicted false positives and actual results were false positives) in the five validation tests were 148, 146, 147, 145, and 147, respectively; the total number of true negatives (model-predicted normal results and actual results were normal) were 125, 123, 125, 125, and 124, respectively; the total number of false positives were 10, 12, 11, 13, and 11, respectively; and the total number of false negatives were 12, 14, 12, 12, and 13, respectively. The average sensitivity was 147.8 / (147.8+12.6) = 92.1% ± 0.8%, the average specificity was 124.4 / (124.4+11.4) = 91.6% ± 0.9%, and the mean AUC was 0.965 ± 0.008 (95% confidence interval: 0.940-0.986). These results demonstrate that the method proposed in this invention exhibits stable and excellent false positive identification performance under different data partitioning conditions. Validation of the practicality of the grading results: In 5 cross-validations, A total of 86 fruits (double-ripened and semi-ripe) were collected. GC-MS analysis showed that approximately 82 of them had an average citral content of less than 45% and a glycoside content of more than 30%, with a positive predictive value of 95.3% ± 1.1%. The average number of fruits was approximately 78 (slightly semi-ripe), of which approximately 69 fruits had a citral content ranging from 45% to 60% and a glycoside content ranging from 18% to 30%, with a positive predictive value of 88.5% ± 1.5%. A total of 131 fruits were tested (normal), of which approximately 124 had an average citral content higher than 60% and a glycoside content lower than 18%, with a negative predictive value of 94.7% ± 0.8%. The above grading results can be directly used to guide the diversion decision for subsequent selective processing.

[0073] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for identifying and grading stress-induced false-ripening fruits of Litsea cubeba, characterized in that, Includes the following steps: S1: Acquire thermal infrared time-series images and hyperspectral images of the target Litsea cubeba fruit; extract the surface temperature change sequence of each fruit from the thermal infrared time-series images; and calculate the thermal time constant. and thermal hysteresis index The thermal hysteresis index Defined as the cumulative difference between the fruit cooling rate curve and the normal reference curve, where the differential region is indicated by a threshold of the second derivative of the rate of temperature change; the epidermal reflectance is obtained from the hyperspectral image through a two-layer radiative transfer model. and the reflectivity of the pulp layer Calculate the pseudo-ripe spectral index at the characteristic wavelength. and glycoside index ; S2: The multiple fruits on the same branch... Using fruit diameter and surface temperature standard deviation as node features, and spatial adjacency relationships between fruits as edges, a spatiotemporal graph sequence is constructed. A spatiotemporal graph convolutional network is then used to extract features from this sequence. This network incorporates an environmental stress memory gating mechanism, which takes temperature, solar radiation, and soil moisture data from the past 72 hours as input and modulates the graph convolution output, ultimately outputting the false ripening probability of each fruit. ; S3: Define the false maturity index and according to the The value is used to grade the fruit: when At that time, it was judged to be a normal fruit. At that time, it was judged to be a lightly ripened, semi-ripe fruit. It was judged to be a heavily overripe fruit; S4: Output the grading results for each fruit, which will be used as the basis for subsequent selective processing decisions.

2. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, Thermal hysteresis index as described in S1 The calculation formula is: in, This represents the number of time sampling points; For the fruit at time Excess temperature; For the normal reference curve at time The excessive temperature, the normal reference curve is established from normal fruit samples that have not been subjected to stress; The curvature threshold of the second derivative; The first derivative of excess temperature with respect to time at time... The value; Number of time sampling points; I This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

3. The method for identifying and grading stress-induced false-ripening fruits of Litsea cubeba according to claim 1, characterized in that, The pseudo-ripe spectral index mentioned in S1 The calculation formula is: in, The reflectance of the pulp layer at a wavelength of 1200 nm; The reflectance of the pulp layer at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1730 nm; The reflectance of the epidermis at a wavelength of 1200 nm; The glycoside index Defined as the ratio of the peak height of the second derivative spectrum to the baseline for hyperspectral reflectance in the range of 1420–1480 nm.

4. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The total reflectivity of the two-layer radiative transfer model described in S1 is expressed as follows: in, Wavelength variable (unit: nm); The reflectance of the epidermis; The transmittance of the epidermis represents the proportion of energy remaining after light passes through the epidermis once. The reflectance of the pulp layer; The absorption and scattering coefficients of the epidermis and pulp layers were fitted using a full-spectrum global parameterization method. The set of global parameters to be fitted is as follows: ， ， ， ， ， ， ， ， ， in, , , These are the concentration weighting coefficients for water, lipids, and citral in the epidermis, respectively. , , These are the concentration weighting coefficients for water, lipids, and citral in the pulp layer, respectively. and These are the scattering coefficients of the epidermis and pulp layer at a wavelength of 1000 nm, respectively. and These are the scattering power indices of the epidermis and pulp, respectively.

5. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The absorption coefficients of the epidermis and pulp layer are respectively expressed as linear combinations of the absorption basis functions of the three chemical components: in, The absorption coefficient of the epidermis; The absorption coefficient of the pulp layer; , , These are the basis functions for the mass absorption coefficients of water, lipids, and citral, respectively. , , , respectively, are the concentration weighting coefficients (unit: g / L) of water, lipids and citral in the epidermis, and are the global parameters to be fitted; , , These are the concentration weighting coefficients (unit: g / L) of water, lipid, and citral in the pulp layer, respectively, and are the global parameters to be fitted. The scattering coefficients of the epidermis and pulp layers are approximated using the power-law approximation of Mie scattering. in, and These are the scattering coefficients of the epidermis and pulp at a reference wavelength of 1000 nm, respectively, and are the global parameters to be fitted. and These are the scattering power indices of the epidermis and pulp, respectively.

6. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The spatial convolution and temporal convolution of the spatiotemporal graph convolutional network described in S2 are defined as follows: in, The node currently being processed; This is the index for the current time slice; For nodes The neighborhood set; These are the normalization coefficients; For nodes In time slice eigenvectors; For nodes In the previous time slot eigenvectors; The weight matrix is ​​the spatial convolution kernel. This is the temporal convolution kernel weight matrix; For nodes In time slice The output feature vector after spatial graph convolution; The implementation method of the environmental stress memory gating is as follows: in, The feature vector output by the spatial graph convolutional layer; The environmental feature vector is represented by principal component scores for each element. To be and The joint vector obtained by concatenation; This is a gated weight matrix, where each element is a weight. This is a bias term, and each element is a bias. It is the Sigmoid activation function. This reduces the gate value to between 0 and 1. This is a gating vector, with each element taking values ​​from 0 to 1, controlling the mixing ratio of new and old features; For Hadamard product; This is the final output feature vector of this node in the previous time slice; This is the final output feature vector after gating modulation, where each element is an activation value.

7. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The method for determining spatial adjacency in S2 is: the distance between any two fruits. Calculated from three-dimensional spatial coordinates: in, and Fruits and fruit Three-dimensional spatial coordinates; For fruit and fruit The Euclidean distance between them. Edge weight The Gaussian kernel function defined as distance: in, This refers to the spatial scale parameter.

8. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The spatiotemporal graph convolutional network comprises three spatiotemporal graph convolutional modules, each outputting the hidden layer feature dimensions. The probability of false maturity The 64-dimensional feature vector output from the third-layer spatiotemporal graph convolution is mapped to a 2-dimensional logical value through a fully connected layer, and then normalized by the Softmax function.

9. The method for identifying and grading stress-induced false-ripe fruits of Litsea cubeba according to claim 1, characterized in that, The training of the spatiotemporal graph convolutional network adopts a two-stage strategy: the first stage uses labeled source domain data for supervised training, employing a binary cross-entropy loss function. in, For the first The true label of each sample (1 indicates false maturity, 0 indicates normal); The model predicts the first The probability of false maturity for each sample; This represents the total number of labeled samples. This represents the cross-entropy loss value. The second stage uses CORAL domain adaptation to fit the unlabeled target domain data, with the loss function being: in, Let be the source domain feature covariance matrix, whose elements are the covariance of the source domain samples on the output features of the third-layer spatiotemporal graph convolution; Let be the target domain feature covariance matrix, whose elements are the covariance of the target domain samples on the output features of the third-layer spatiotemporal graph convolution; For feature dimensions; It is the Frobenius norm.

Images